R doParallel foreach with error handling for independant workers

Question

I have to run a lot of random forest models so I want to use doParallel on my server with 8 cores to speed up the process.

Yet, some models need a lot longer than others or even might throw errors. I would like to run 8 models in parallel, and if a model throws an error and/or is skipped then the workers should just continue. Each model result is saved on harddrive so I can access and combine them later.

TryCatch

or

.errorhandling="remove"

did not solve the problem. I get

 Error in unserialize(socklist[[n]]) : error reading from connection

Code example: I tried it with %do% and model 2-7 run successfully. Yet in %dopar% I get the shown error

 foreach(model=1:8, .errorhandling="remove") %dopar% {


      tryCatch({
          outl <- rf_perform(...)
          saveRDS(outl,file=getwd() %+% "/temp/result_" %+% model %+% ".rds")

     }, error = function(e) {print(e)}, finally = {})
  }

note: i am running newest R 3.2.3 on Ubuntu 14.04 with 16 cores and 64G ram — user670186, Dec 29 '15 at 16:05

score 3 · Accepted Answer · edited Mar 24 '17 at 12:56

I think I found the problem: if the objects you export to the clusters are too big, either R cant handle it anymore and/or there is a timeout

My data object exportet was 5 million rows and 300 variables, that exported to 16 clusters.

cl <- makeCluster(16)
registerDoParallel(cl)
clusterExport(cl, "data")

#data must not be too large

I downsized the object into smaller pieces and now it works. The authors might want to mention that in the documentation of doParallel or throw a warning if objects are too big.

R doParallel foreach with error handling for independant workers

1 Answers1